apprun.icuGuides

Guide · Creative AI

The anatomy of an AI image prompt: subject, style, light, lens and parameters

A prompt is a set of parts, and each part controls one thing. Name the subject, setting, camera, light, style and settings, see which are missing, and keep them separate so you can change one at a time.

By apprun.icuUpdated Read 9 minIncludes a free browser toolSources 8 cited
On this page
  1. The short answer
  2. What are the parts of an AI image prompt?
  3. How do you dissect a real prompt?
  4. What are the six types of prompt modifier?
  5. Why does the same prompt behave differently in each tool?
  6. Are quality words worth adding?
  7. How does the tool colour-code the parts?
  8. Common mistakes and what this cannot tell you
  9. Common questions
  10. Sources and further reading

What are the parts of an AI image prompt?

This guide splits an image prompt into eleven parts. You need not fill them all, but a part you leave out is left to the model's defaults. The table gives each part in plain language, with an example phrase.

The eleven parts of an image prompt. The examples are ours, not taken from a source.
PartWhat it controlsExample phrase
SubjectThe main thing shownan old fishing boat
ActionWhat the subject is doingdrifting out of the harbour
SettingWhere the scene happenson a grey harbour
Composition and cameraFraming, viewpoint and lenslow camera angle, 35mm
LightingLight source, time of day or qualitysoft morning light, golden hour
Style and mediumThe rendering or the mediumphotograph, watercolour
ColourThe palettemuted blue and rust
MoodThe feeling the image should givequiet, lonely
Quality wordsWords meant to add detail or polishhighly detailed, 8k
Negative promptWhat to avoid, in a separate field where a tool has onepeople, text, watermark
ParametersSettings kept apart from the sentenceaspect ratio 3:2
One prompt becomes labelled parts, then a line you can reassemble in orderYour promptone lineof wordsLabel partssubject, settingcamera, lightstyle, colourFind gapsmissing partsand conflictsReassemblein a fixed ordersettings kept apart
Each part is named before it is changed, and settings stay outside the sentence.

How do you dissect a real prompt?

Take one prompt and see how the tool splits it. The example is invented, and the table shows what the tool returned in this version.

An old fishing boat on a grey harbour at dawn, seen from a low camera angle on the dock, soft fog, muted blue and rust colours, quiet mood, highly detailed, 35mm photograph
Negative prompt: people, text, watermark
Aspect ratio: 3:2
What the tool returned for the example text. The tool may change its rules, so check its output for your own text.
FieldWhat the tool found
DialectWeighted-tag, because of the Negative prompt line
Subject"An old fishing boat"
ActionNone
Setting"at dawn"
Camera"low camera angle", "35mm"
LightingNone, so the tool reports a gap
Style"photograph"
Colour"muted"
Mood"quiet mood"
Quality words"highly detailed"
Detail"on a grey harbour", "seen from a", "on the dock, soft fog", "blue and rust colours"
Negative prompt"people, text, watermark"
Parameters"Aspect ratio: 3:2"

The tool also reports three checks: a gap for missing lighting words, a note that a negative prompt may be ignored by some models, and a note that quality words are not documented as switches. Its detail spans are the fragments left between recognised phrases, so "seen from a" stands alone. Reassembled in the tool's fixed order, the example reads:

An old fishing boat, at dawn, low camera angle, 35mm, photograph, muted, quiet mood, highly detailed, on a grey harbour, seen from a, on the dock, soft fog, blue and rust colours
Negative prompt: people, text, watermark
Aspect ratio: 3:2

To check your own prompt, paste it into Image Prompt DNA.

What are the six types of prompt modifier?

Oppenlaender's taxonomy, arXiv 2204.13988 (version 3, revised 14 June 2023), names six types of modifier used by practitioners.

Six modifier types. Definitions are quoted from the paper's Table 1; examples are ours.
TypeDefinition in the paperExample (ours)
Subject term"Denotes the subject"fishing boat
Style modifier"Indicates an artistic style"watercolour
Image prompt"Indicates a style or subject via an image"a reference picture
Quality booster"A term intended to improve the quality of the image"highly detailed
Repeating term"Repetition of subject terms or style terms with the intention of strengthening this subject or style"fishing boat, old fishing boat
Magic term"A term that is semantically different from the rest of the prompt with the intention to produce surprising results"an unrelated noun added to a harbour scene

Modifiers can overlap, because "Modifiers either modify the image style or enhance its quality, and these two types can overlap". The paper also calls repeating terms "solidifiers", and says they "can be applied to any of the other types of modifiers". Treat the six types as vocabulary, not evidence: the paper reports no measurement of what any modifier does to an image.

Why does the same prompt behave differently in each tool?

Each tool family reads its own syntax, and some have no field for a part at all. A blank cell below means the page we read did not cover that point.

Syntax by tool family, as the sources read on 17 October 2026.
Tool familyWeightNegative prompt
AUTOMATIC1111 web UI (wiki, last edit 15 December 2024)"(word:1.5)" raises attention by a factor of 1.5; "(word)" is 1.1A separate field for "things the model should avoid"
FLUX 3 Image (Black Forest Labs)Not covered"FLUX 3 Image has no negative prompt field."
FLUX.2 [pro] and [max]Not covered"No negative prompts supported"
GPT Image (OpenAI, no date shown)Not coveredNot mentioned
Imagen on Vertex AI (Google, no longer updated)Not coveredGiven "along with the original prompt"

OpenAI's image guides name quality, size, format, compression, n and partial_images as settings, and say to "Set API parameters separately from the prompt." The FLUX 3 quick reference says "Unknown fields return 422."

Weight syntax is the least portable part. The AUTOMATIC1111 wiki says NovelAI, which it abbreviates as NAI, uses a 1.05 multiplier and curly braces, so a weight copied between tools does not keep its meaning. This guide does not describe flag syntax such as --ar, because the Midjourney documentation could not be read.

Check dates before you copy advice. The OpenAI pages show no publication date, and the Google guide carries a banner that says "Vertex AI documentation is no longer being updated".

Are quality words worth adding?

Quality words such as "highly detailed" or "8k" are worth adding only as a test, because none of the sources we read shows one working as a switch for quality. The taxonomy paper says quality boosters can "increase aesthetic qualities and the level of detail in images", and that verbosity "may boost the amount of details and overall quality of the generated image, at the expense of the subject becoming less controllable". It reports no measurement of either effect. For its FLUX 3 Image model, Black Forest Labs tells users to "omit quality tags such as “masterpiece”."

Liu and Chilton's VQGAN+CLIP test did not look at quality words. It did look at connecting words, and found that they "did not contribute statistically meaningful differences in generation quality". That tested an older model, so test a quality word yourself: generate the same subject and settings with and without it, and keep it only if the result meets a standard you wrote down first.

How does the tool colour-code the parts?

The tool marks each part where it finds the words for it, so you can see which words it read as subject, lighting or style. Each part lists its values, the exact span of text they came from and any weight you typed. Parameters and negative prompts are kept as written.

The Image Prompt DNA tool also flags gaps, which are parts with no words. It checks gaps for subject, camera, lighting and style. It also flags conflicts, such as photograph with oil painting, macro with wide shot, or golden hour with a city street at night. A weight above 1.5 is flagged as a possible issue, not an error. Gaps and conflicts are questions to answer, not verdicts: a conflict may be the look you want. Once a prompt is labelled, vary one part at a time with the guide to controlled image prompt variations.

Common mistakes and what this cannot tell you

  • Treating a quality word as a switch. Test it against the same prompt without it.
  • Copying a weight from one tool to another. NovelAI's convention differs.
  • Writing "no" or "don't" in a negative field. An Imagen guide that Google marks as no longer updated advises against it: "Avoid instructive language or words like “no” or “don’t”." Black Forest Labs suggests describing the scene instead, as "an empty promenade" instead of "no crowds".
  • Expecting a negative prompt everywhere. FLUX 3 Image has none.
  • Reading the tool's labels as proof of how an image will look. The tool checks wording only.

This guide also does not describe Midjourney, whose documentation could not be read during our research.

Common questions

Do I need all eleven parts in an image prompt?

No. The eleven parts are a checklist, not a requirement. Name the parts you care about and decide on purpose what to leave out. A part you leave out is left to the model's defaults, so a prompt with no lighting words gets whatever light the model chooses. The tool's gap check covers subject, camera, lighting and style.

What is a negative prompt, and does every tool have one?

A negative prompt lists what the image should avoid. Not every tool has one. The AUTOMATIC1111 wiki offers a separate field for "things the model should avoid". Black Forest Labs says FLUX 3 Image has no negative prompt field, and that FLUX.2 [pro] and [max] support none. Check your tool's documentation before you rely on the field.

Does the same prompt give the same image in every tool?

No. Tools differ in weight syntax, in whether they have a negative field and in which settings sit outside the words. None of the sources we read shows a prompt carrying over unchanged between tools. Expect to rewrite the parts for each tool, then test the result the way you would test any change.

Are quality words such as masterpiece or 8k worth adding?

Test them rather than assume them. Black Forest Labs tells FLUX 3 Image users to omit quality tags such as "masterpiece", and the taxonomy paper describes quality boosters only by their intent. Generate the same subject with the same settings, once with the words and once without, and keep them only if the result meets the standard you wrote down before you generated.

What does (word:1.5) mean in an AI image prompt?

In the AUTOMATIC1111 web UI syntax, (word:1.5) increases the attention given to that word by a factor of 1.5. A plain (word) is a factor of 1.1, and square brackets lower attention. The wiki says NovelAI uses a 1.05 multiplier and curly braces, so a weight from one tool does not carry over to another.

Sources and further reading

Share this guide